JP2001109844A - Character string extracting method, handwritten character extracting method, character string extraction device, and image processor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To extract the character string of a handwritten address, etc., fast with high precision while evading a complicated integrating process in all pixels. SOLUTION: A connecting pixel group (CC) detection part 13 detects a connecting pixel group (CC) formed of black connecting pixels in a binarized image, a character size connecting pixel group (CharCC) extraction part 14 extracts a character size connecting pixel group (CharCC) which is neither too large nor too small among detected connecting pixel groups, and a lateral expansion part 21 and a longitudinal expansion part 22 which expand and contract the extracted character size connecting pixel group by expanding it in a virtual character string direction and contracting it at right angles to the virtual character string direction. Then thin and long connecting pixel group (LongCC) extraction parts 25 and 26 combine expanded and contracted connecting pixel groups in the virtual character string direction to extract thin and long connecting pixel groups and a character string selection part 30 selects an object character string of image recognition according to the extracted long and thin connecting pixel groups.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image processing apparatus and the like used for, for example, an automatic address reader for mail, and more particularly, to an image processing apparatus suitable for extracting an inconsistent character string such as Japanese handwritten address characters. The present invention relates to a processing device and the like.

[0002]

2. Description of the Related Art For example, in the field of mail processing, it is necessary to process a large number of mails every day in a short time. In the past, humans were classified visually and manually, and the burden on postal workers was very large. To this end, the mechanization with automated reading devices has been promoted,
For example, great progress has been made in the classification of each service area, such as reading postal codes written in the postal code column in Japan. However, it is difficult to recognize the position of the postal code for postal items from other countries without a postal code column. Also, in the future mail processing, which is expected to increase more and more, the development of technology to immediately read the address information on the mail other than the zip code and classify it into service areas,
It is a technical matter that is strongly desired in the future.

As described above, although there is a strong demand for extracting a postal code from a mail having no postal code column and reading address information, it is very difficult to read such information from a mail using current technology. Difficult. this is,
In addition to the address area, the mail area includes the sender's area, advertising information,
The main reason is that various patterns such as pictures and photographs are included, and there are too many variations such as a mixture of handwritten and printed characters, vertical writing and horizontal writing.
Therefore, in order to correctly extract postal code and address information from such mixed information, very complicated processing such as searching all areas is required. In particular, in handwritten addresses in Japanese, in addition to the problem of mixed vertical and horizontal writing, handwritten characters forming the addresses are often arranged irregularly, and it is particularly difficult to perform good address reading.

Here, as an existing method for extracting address information, there is a technique of first extracting an address area and then reading an actual address. For example, assuming that the print address is attached to a postal matter as a label, the method of extracting the address area by detecting its reflection attribute and the shadow of the edge of the label, or projecting in the horizontal and vertical directions There is a method of estimating a destination area by using a high density place as a character line. Japanese Patent Application Laid-Open No. Hei 7-265807 discloses a technique of estimating a character line by integrating adjacent pixel connected components, and further integrating them to be a candidate for a destination area.

[0005]

However, in the method of extracting an area on the assumption that the print address is attached as a label, the label is not attached and the print address is directly written on the mail. Cannot be handled at all, and the mail that can be extracted is greatly limited. In particular, it is difficult to apply this technique to handwritten Japanese addresses where labels are rarely used. Also, in the method of estimating a place with high density as a character line, it is difficult to distinguish a mail having a complicated pattern with a lot of image information in a background portion from an area other than the address area, so that accurate addressing is not possible. The area cannot be detected. Further, JP-A-7-26
The technology disclosed in Japanese Patent No. 5807 assumes that characters are relatively neatly arranged and character strings are relatively densely arranged. When applied to regular handwritten addresses, satisfactory results cannot be expected. Further, in this technique, it is necessary to perform a complicated integration process in all pixels, such as picking up a cluster from all the pixels and repeating the operation of searching for all the clusters for each cluster. Therefore, the logic is very difficult, the processing speed is extremely slow, and the configuration is too complicated, so that it cannot be said to be realistic from the viewpoint of the system.

SUMMARY OF THE INVENTION The present invention has been made to solve such a problem, and it is an object of the present invention to accurately and quickly extract a character string such as a handwritten address by avoiding complicated integration processing in all pixels. Main purpose. It is another object of the present invention to provide a character string extraction that can be put to practical use even if the arrangement of the character strings is irregular. Still other purposes are vertical writing,
It is an object of the present invention to extract character strings by providing flexibility and easy processing even for an image of a postal matter or the like in which horizontal writing is mixed. Still another object is to provide a high flexibility by processing an algorithm for a character string matched with a print character string or the like and an irregular character string such as a Japanese handwritten address. And highly accurate character string extraction.

[0007]

With this object in mind, the character string extracting method according to the present invention comprises the steps of: extracting a connected pixel group of pixels constituting an input image; and extracting the connected pixel group of the extracted connected pixel group. Extracting a connected pixel group within the threshold by comparing the size with a predetermined threshold, and extracting a connected elongated pixel group extending in the vertical or horizontal direction of the extracted connected pixel group within the threshold. And a step of extracting a character string candidate based on the connected state of the extracted elongated connected pixel group.

As a step of extracting the connected pixel group, a method of extracting the connected pixel group by using an eight-connected pixel group vertically, horizontally, and diagonally or a four-connected pixel group vertically, horizontally, and extracting a pixel which falls within a predetermined threshold value is used. Is mentioned. In addition, the step of extracting the connected pixel group within the threshold value is performed in a vertical direction of the connected pixel group and / or
Alternatively, the intra-threshold connected pixel group may be extracted by comparing the horizontal direction with predetermined vertical and / or horizontal directions of the assumed character size. According to this, for example, by determining a threshold value in consideration of a character size expected to be a handwritten character and a character size, it is possible to detect a connected pixel group in a state where noise is removed from a postal matter in which patterns and the like are mixed. It becomes possible. More specifically,
For example, if the height or width of the connected pixel group is equal to or smaller than a first constant value (for example, 110 pixels at a resolution of 4 lines / mm), the connected pixel group is set to a threshold connected pixel group.

Further, the step of extracting the elongated connected pixel group is characterized in that the elongated connected pixel group is extracted by changing the extension size based on the size of the extracted intra-threshold connected pixel group. This is preferable because it can correspond to the size of handwritten characters with many variations. More specifically, if the length of the in-threshold connected pixel group in the extending direction is L, it may be extended on both sides by L. Further, the step of extracting the elongated connected pixel group is configured to extract the elongated connected pixel group by contracting the horizontal direction when the vertical direction is extended and contracting the vertical direction when the horizontal direction is extended. For example, it is excellent in that a character string can be extracted while preventing contact between adjacent characters. Moreover,
The step of extracting the elongated connected pixel group is characterized in that the elongated connected pixel group is extracted by changing the amount of contraction based on the size of the extracted connected pixel group within the threshold, as described above. This is preferable in that it can correspond to the size of handwritten characters having many variations.

Further, assuming a mesh image obtained by dividing the image into a predetermined size, turning on a mesh image corresponding to the extracted connected pixel group within the threshold value, and determining a connected state of the mesh images turned on. Extracting a specific region.
Further, if the image is a print character, the process proceeds to a step of turning on the mesh image. If the image is a handwritten character, the process proceeds to a step of extracting a group of elongated connected pixels. This feature is advantageous in that a destination area can be extracted by a simpler area extraction method for print characters, and the processing can be speeded up. Note that the method of extracting a specific region from the mesh image is a method of shifting from the world of one pixel per pixel to the world of the mesh image which is a predetermined rectangular area, and simplifying this mesh image as one point. It is possible to extract a specific area with a simpler algorithm.

In the character string extracting method according to the present invention, a pixel group having a predetermined size is extracted from an image including an input handwritten character, and the pixel group is expanded by extending one direction and contracting the other direction. Character string in the direction of the character string, detecting a string state of the assumed string of plural strings, and estimating an address pattern from the detected string state of the string. . Here, the detected string state of the character string includes the string length and the unevenness state of the string, and the order in which the character recognition processing is performed is that of the long string length and / or the unevenness of the string. It is preferable to increase the order of the items having a small number of states, since a more appropriate character string can be extracted as a character string candidate for an address or the like. In addition, the string of the assumed character string is extended by adding a predetermined ratio to the size of this pixel group with respect to the vertical or horizontal one of the pixel groups, and is expanded with respect to the other horizontal or vertical. A character string may be assumed based on a connection state of a plurality of pixel groups that are reduced, expanded, and reduced by reducing a predetermined ratio to the size of the pixel group.

Further, the character string extracting apparatus according to the present invention includes a connected pixel group detecting section for detecting a connected pixel group formed from black connected pixels in a binarized image, and a connected pixel group detecting section. Among the detected connected pixel groups, a character size connected pixel group extraction unit that extracts a character size connected pixel group that is neither too large nor too small, and a character size connected pixel group extracted by the character size connected pixel group extraction unit. On the other hand, a connected pixel group expansion / contraction unit that expands / contracts the character size connection pixel group by extending in the assumed character string direction and contracting in the direction orthogonal to the assumed character string direction, An elongated connected pixel group extraction unit for connecting the connected pixel groups in the assumed character string direction to extract an elongated connected pixel group, and image recognition based on the elongated connected pixel group extracted by the elongated connected pixel group extraction unit. It is characterized in that a character string selection unit for selecting a character string to. Further, the character size connected pixel group extracting unit extracts the character size connected pixel group based on a threshold value in consideration of the size of the handwritten character written on the mail, and the character size connected pixel group This is preferable in terms of extracting columns. Here, the assumed character string direction is a vertical direction when considering a character string of a vertically written address, and a horizontal direction when considering a character string of a horizontally written address.

The image processing apparatus according to the present invention comprises: an image input means for inputting an image; a pixel group extracting means for extracting a pixel group having a predetermined size from the image input by the image input means; A vertical extension that defines a vertical direction and a horizontal direction with respect to the pixel group extracted by the pixel group extraction means, and generates a vertical extension pixel group in which the pixel group is extended and contracted in the horizontal direction. A pixel group generating means and a vertical direction and a horizontal direction are defined with respect to the pixel group extracted by the pixel group extracting means, and in this horizontal direction, the pixel group is expanded and the horizontal direction is reduced in the vertical direction. A horizontally elongated pixel group generating means for generating a pixel group; a plurality of vertically elongated pixel groups generated by the vertically expanded pixel group generating means and a plurality of horizontally expanded pixels generated by the horizontally expanded pixel group generating means Group Elongated connected pixel group extracting means for extracting elongated connected pixel groups by combining in the extended directions, and extracting a character string candidate based on the elongated connected pixel group extracted by the elongated connected pixel group extracting means. It is characterized by comprising a character string candidate extracting means and an image recognizing means for recognizing an image with respect to the character string candidate extracted by the character string candidate extracting means.

The character string candidate extracting means outputs a plurality of elongated connected pixel groups extracted by the elongated connected pixel group extracting means with higher priority to those having a longer length. can do. Further, the character string candidate extracting means determines a priority by lowering an evaluation point with respect to a pixel group intersecting the vertical direction expanded pixel group and the horizontal direction expanded pixel group. This is preferable in that column extraction can be prevented and an image processing apparatus which is more practical can be provided.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail based on an embodiment shown in the accompanying drawings. FIG. 1 is an explanatory diagram illustrating the overall configuration of the image processing apparatus according to the present embodiment. As shown in the figure, the image processing apparatus according to the present embodiment is largely composed of a character string extraction device 1, an image input device 2, and an image recognition device 3. The image input device 2 includes, for example, a scanner that has a line sensor in the main scanning direction and scans an image by scanning in the sub-scanning direction, and an optical character reading device that reads an entire image by a photo. It is configured so that an entire image can be obtained. Also, the image input device 2
Is provided with a photoelectric conversion circuit (not shown), and is configured to convert an optically read image into an electric signal.

Next, the character string extracting device 1 according to the present embodiment will be described. The electric signal output from the image input device 2 is converted by the binarization unit 11 into two signals of a white pixel and a black pixel.
It is converted to a value signal. The binarizing unit 11 also performs noise removal for removing small noise. In some cases, the image input device 2 includes a binarizing unit.
The value may be input to the character string extraction device 1 in some cases. In this case, the character string extracting device 1 will not have the binarizing unit 11. Next, the binarized image data is reduced by the reduction unit 12. For example, in the reduction unit 12, an image having a resolution of 8 lines / mm
It is configured to convert to a resolution of book / mm, and a reduction operation is performed to increase the processing speed. As a simple reduction method, for example, there is a method of performing an AND operation on continuous lines and outputting the result. However, since the accuracy of region extraction is affected, if there is no problem in the processing speed, the reduction unit 12 is used. No need to provide. It is also possible to reverse the order of the binarizing unit 11 and the reducing unit 12.

The image data reduced from the reduction unit 12 is input to a connected pixel group (CC) detection unit 13. In the connected pixel group (CC) detection unit 13 in the present embodiment, an eight connected pixel group, that is, a connected pixel group (Connected Component: C
C) is all extracted. The eight connected pixel group is for detecting a connected pixel group (CC) by capturing eight pixels that are left, right, up, down, and oblique with respect to the center pixel, and extracting the connection of these black pixels. As a method of detecting the connected pixel group (CC), there is also a method of extracting four connected pixel groups, that is, left, right, and up and down. From experience, it is preferable to employ an 8-connected pixel group.

The threshold value storage unit 15 stores a threshold value of a connected pixel group (CC) used for determining a character size.
In the character size connected pixel group (CharCC) extraction unit 14, the connected pixel group detected by the connected pixel group (CC) detection unit 13
(CC) is compared with the threshold value stored in the threshold value storage unit 15, and a character size connected pixel group (Cha
rCC). First, the threshold storage unit 15 stores a maximum constant value assuming that the height or width of the connected pixel group (CC) does not exceed a predetermined amount, and stores a threshold value such as 110, for example. . This value is
It is provided to remove large patterns and long ruled lines in printed mail, etc., and has 110 pixels at a resolution of 4 lines / mm, that is, a connected pixel group exceeding about 27 mm.
(CC) is defined so as to be excluded from the character size connected pixel group (CharCC). Also, as a condition that the rectangular area (rectangular area) of the connected pixel group (CC) does not take a value smaller than a predetermined size, a value that the rectangular area is 3 or more and the rectangular area is 8 or more is also stored. I have. This value is determined so as to remove a rectangular area of 3 pixels × 8 pixels or less in order to remove a thin line from a noise, such as an edge in a label, and to remove a broken line.

On the other hand, the character size connected pixel group (CharCC) extracted by the character size connected pixel group (CharCC)
CC) is input to the size determination unit 16. In the size determination unit 16, for example, the pitch of a large character size connected pixel group (CharCC) is measured, and the larger the pitch, the higher the possibility of a handwritten character, and the smaller the pitch, the higher the possibility of a printed character. Is determined. This is because, in general, the pitch between characters varies in the case of handwritten characters, and if variations occur, it is determined that the possibility of a printed character is low. When it is determined that the possibility of a handwritten character is high based on the result of the size determination unit 16, a character string is extracted by the handwritten character string extraction unit 20, and it is determined that the possibility of a print character is high. In this case, a character string is extracted by the print character string extracting unit 40. Note that the determination by the size determination unit 16 is not limited to the above method.
It is also possible to configure so that the determination can be made by an arbitrary method, such as by measuring the frequency distribution of (CharCC), or by a combination thereof.

Next, the configuration of the handwritten character string extracting unit 20 in the present embodiment will be described. Size determination unit 16
If it is determined that the possibility of handwritten characters is high,
Assuming that the character string direction is the vertical direction and the horizontal direction, each processing is executed independently. First, in the case of a process assuming that the character string direction is the horizontal direction, the horizontal direction expansion unit 21 expands the extracted character size concatenated pixel group (CharCC) in the horizontal direction that is the assumed character string direction, Processing for contracting in the vertical direction which is a direction orthogonal to the assumed character string direction is performed. This expanded and contracted character size connected pixel group (CharCC) is Hor
The data is stored in the first buffer 23. Also, the elongated connected pixel group
The (LongCC) extraction unit 25 combines the expanded and contracted character size connected pixel groups (CharCC) to extract a horizontally elongated connected pixel group (LongCC). The horizontal evaluation unit 27 evaluates the connected state of the extracted elongated connected pixel group (LongCC),
The evaluation result is output to the character string selection unit 30.

On the other hand, in the case of processing assuming that the character string direction is the vertical direction, the vertical extension unit 22 applies the extracted character size concatenated pixel group (CharCC) to the vertical direction which is the assumed character string direction. , And shrink in the horizontal direction, which is a direction orthogonal to the assumed character string direction. The expanded and contracted character size connected pixel group (CharCC) is stored in the Ver first buffer 24. In addition, the elongated connected pixel group (Long C
C) The extraction unit 26 combines the expanded and contracted character size connected pixel groups (CharCC) to form a vertically elongated connected pixel group (L
ongCC). The vertical evaluation unit 28 evaluates the connected state of the extracted elongated connected pixel group (LongCC) and outputs the evaluation result to the character string selecting unit 30.

The character string selection unit 30 mixes the elongated connected pixel group (LongCC) stored in the Hor first buffer 23 and the Ver first buffer 24 with the horizontal evaluation unit 27 and the vertical evaluation unit. 28 are sorted in descending order of evaluation score. In addition, the character string threshold storage unit 2
Based on the threshold information stored in No. 9, if the evaluation points are equal to or higher than the threshold in the sorted order, the character string is determined.
Here, in the present embodiment, as described later, when a certain horizontal elongated connected pixel group (LongCC) is selected as a character string, this horizontal elongated connected pixel group (LongCC) The character string of the direction is configured to be used as a deduction material when subsequently selected. Therefore, the elongated connected pixel group
If (LongCC) is the horizontal direction, the orthogonal deduction points for Ver
The second buffer 32 stores deduction point rectangular area information (circumscribed rectangle of LongCC). Similarly, the elongated connected pixel group (Lo
If (ngCC) is the vertical direction, the deduction point rectangular area information is stored in the orthogonal second point deduction Hor second buffer 31.
The character string selecting unit 30 sequentially sorts and sorts the character string candidates while sequentially reducing the evaluation points, and sends the character string information to the image recognition device 3 in descending order of the evaluation points. . In the image recognition device 3, a plurality of address candidate areas are normally extracted based on the binarized signal of the entire image from the binarizing unit 11 and the character string information from the character string selecting unit 30. The address reading is performed in order from the area having the highest evaluation score and the highest priority, and the reading operation is completed when a satisfactory result is obtained.

Next, the configuration of the print character string extracting unit 40 in the present embodiment will be described. The print character string extraction unit 40 processes characters having a relatively small character size concatenated pixel group (CharCC), for example, 30 pixels at a resolution of 4 lines / mm as a threshold, that is, characters of about 8 mm or less. First, the mesh image forming unit 41 forms a so-called mesh image that divides an image into small areas and assumes a simplified rectangular area as one point. This mesh division method is effective, for example, to determine the size of the mesh in a range equivalent to the large character size from the assumed address, for example, the largest character size connected pixel group (CharCC) Is 30 pixels, the mesh is divided into meshes in a rectangular area consisting of a small area of 24 pixels × 24 pixels, which is a predetermined size.

Next, in the corresponding mesh image detecting section 42,
The character size connected pixel group (CharCC), which is the extracted intra-threshold connected pixel group, detects which (corresponds to) which mesh image is a rectangular area assumed by the mesh image forming unit 41. , The corresponding rectangular area is selected and the mesh image is turned ON. As a method of this detection, for example, an extracted character size connected pixel group (CharC
A method of turning on the mesh image to which the center coordinate of C) belongs is exemplified. Further, when a small mesh is assumed by the mesh image forming unit 41, there is a method of turning on a mesh image in which the rectangles of the character size connected pixel group (CharCC) slightly overlap. Upon receiving the output from the corresponding mesh image detection unit 42, the mesh image connection detection unit 43 detects the connection state of the meshes that have been turned ON. For example, in order to make the same area even if separated by one line, one unit of mesh sandwiched between meshes that are turned on in the vertical and horizontal directions is turned on. Also, in the mesh image, a 4-connected mesh in which the mesh is connected vertically and horizontally or an 8-connected connected mesh (Conne
cted Mesh: CM).

Upon receiving the result from the mesh image connection detecting section 43, the specific area extracting section 45 extracts a destination area candidate in consideration of the attribute information from the attribute information storing section 44. The attribute information storage unit 44 stores threshold information for selecting a destination area candidate whose size is neither too small nor too large. Also, for example, 2
There is a character string of more than lines, or a sufficient number of connected pixel groups (CC)
Is stored, and attribute information such as whether or not the character string is in the form is stored. In addition, since the address information of a mail is usually shifted to the center of the mail, attribute information such as whether or not the area is closer to the end of the entire image or whether the area is closer to the center is also stored. I have. In the specific area extraction unit 45,
Based on the attribute information from the attribute information storage unit 44, for example, the connected mesh by the mesh image connection detection unit 43
By assigning an evaluation point to (CM), a specific area that is a destination area candidate is extracted. As a result, the address area can be limited to several, and by outputting this result to the image recognition device 3 that actually recognizes the address, the address reading accuracy can be remarkably improved and the address reading can be performed. Can be speeded up. Print character string extraction unit 40
In the image recognition device 3 which is a post-processing device, the binarizing unit 11
For example, based on the binarized signal of the whole image from and the area information from the specific area extraction unit 45, the address reading is performed in order from the area having the highest evaluation point and the highest priority, and a satisfactory result is obtained. The reading operation is completed at the stage.

As described above, the print character string extraction unit 40 according to the present embodiment shifts the world from a pixel unit, which is one pixel per pixel, to a mesh world, which is an area having a predetermined size. Since the simplification method is adopted as one point, the address area can be extracted at a higher speed with a simpler algorithm than the existing method. Furthermore, highly accurate area extraction is realized by performing extraction in consideration of the character size of the address, the connectivity of the mesh, attribute information, and the like. In addition, by performing the address reading from an area having a high priority with an evaluation point assigned, useless address reading work can be minimized and the address reading can be performed more efficiently and quickly.

Next, the flow of a mail address reading operation according to the present embodiment will be described with reference to FIGS. Here, FIG. 2 shows a rough flow of address reading mainly with a handwritten character string in the present embodiment. FIGS. 3 to 8 show a first practical example as an original image.
(Figure 3), extraction state of character size connected pixel group (CharCC)
(FIG. 4), the extraction state of the horizontal elongated connected pixel group (horizontal LongCC) (FIG. 5), the extraction state of the vertical elongated connected pixel group (vertical LongCC) (FIG. 6), and the character strings are selected and ranked. Attached
FIG. 7 shows an extracted character string superimposed on the original image (FIG. 8).

First, an image of a mail is read by the image input device 2 (step 101). The postal matter 50 as shown in FIG. 3 includes, in addition to the address information 51, various kinds of image information such as a postal code 52, distribution information 53, sender information 54, a picture 55, and the like. These image information are stored in a binarizing unit 1
1 is binarized by an appropriate method (step 102),
The connected pixel group (CC) is detected from the binarized image by the connected pixel group (CC) detection unit 13 and registered in a memory (not shown) (step 103). This connected pixel group (CC)
Is compared with the threshold value stored in the threshold value storage unit 15 by the character size connected pixel group (CharCC) extraction unit 14, and a character size connected pixel group (CharCC) that is neither too large nor too small is extracted (step 104). ). FIG. 4 shows a state in which the character size connected pixel group (CharCC) 56 is formed into a black rectangle and is superimposed on the original image for easy understanding. As shown in FIG. 4, almost all characters are extracted as a character size connected pixel group (CharCC) 56 of various sizes. However,
The characters “•” (dot) and “2” are not extracted because they do not satisfy the threshold condition described above. Also, picture 55
It can be understood that many parts constituting are not similarly extracted.

Next, assuming that the character string direction is the vertical direction and the horizontal direction, the following processing is executed independently (step 10).
5). First, a horizontal extension unit 2 as a connected pixel group expansion / contraction unit
1 and the vertical extension unit 22 form a rectangle surrounding each character size concatenated pixel group (CharCC), the longer one of the length and width is L, and the rectangle is assumed to be the character string direction (vertical or horizontal).
Is extended by L on each side, and is reduced by, for example, (2/3) · L in the direction orthogonal to the assumed character string direction (step 10).
6) In the horizontal direction and the vertical direction, the elongated connected pixel group (Long
CC) is formed.

FIG. 5 is a diagram showing each character size connected pixel group (CharC
FIG. 11 is a diagram in which a rectangle surrounding C) is extended in the horizontal direction and contracted in the vertical direction, and is superimposed on the original image for easy understanding. However, the character image of the original image is not used for the following character string selection processing. As shown in FIG. 5, the circumscribed rectangle of the character size connected pixel group (CharCC) is expanded in the horizontal direction, so that the horizontal elongated connected pixel group connected in the horizontal direction is formed.
(LongCC) 57 is formed. Also, by reducing the size in the vertical direction, contact between adjacent character strings is prevented. Also,
FIG. 6 is a diagram in which a rectangle surrounding each character size concatenated pixel group (CharCC) is extended in the vertical direction and contracted in the horizontal direction, and is superimposed on the original image for easy understanding. As shown in the figure, the longer one of the circumscribed rectangles of the character size connected pixel group (CharCC) is L, the rectangle is extended by L on both sides in the vertical direction, and the rectangle is reduced in the horizontal direction (for example, (2/3) ) · L) to form a vertically elongated connected pixel group connected in the vertical direction.
(LongCC) 58 is formed. Here, in the present embodiment, the value of expansion and contraction of the circumscribed rectangle is not a fixed value, but is determined from the size (L) of the character size connected pixel group (CharCC). This is because the variation in character (especially handwriting) size is large, and it is not appropriate to use a fixed value, and the character spacing is considered to be almost proportional to the character size.

Next, the elongated connected pixel group (LongCC) extraction units 25 and 26 extract the elongated connected pixel group (LongCC) obtained by combining the character size connected pixel group (CharCC) by the extension in the assumed character string direction. Then, evaluation points are given by the horizontal evaluation unit 27 and the vertical evaluation unit 28 (step 10).
7). For example, the evaluation point is higher as the length is longer in the assumed character string direction.For example, the connected pixel group (CC) is
The less the unevenness in gCC), the higher the evaluation point. However, at this stage, unlike the original character strings, the character groups in the horizontal direction shown in FIG. 3, such as “ta”, “ei”, “government”, and “stock”, are pseudo character strings arranged in the orthogonal direction. Nevertheless, since it is relatively long like the elongated connected pixel group (LongCC) 61 in FIG. 5, a high evaluation point is given.

Next, the character string selection unit 30 includes two buffers (the first Hor buffer 23 and the first Ver buffer 2).
(4) The elongated connected pixel group (LongCC) in (4) is mixed and sorted in descending order of evaluation points (step 108). Then, checking is performed in order from the longest connected pixel group (LongCC) having the highest evaluation point. If the evaluation point is equal to or more than the threshold value stored in the character string threshold value storage unit 29, the character string is determined. Orthogonal elongated connected pixel group that intersects with the determined character string (orthogonal LongC
In C), the evaluation points are deducted according to the intersection area etc.
(Step 109). As described above, for the elongated connected pixel group (LongCC) determined to be a character string, the orthogonal buffer (Ho) is used.
A black rectangle for deduction (circumscribed rectangle of LongCC) is written in the r second buffer 31 and the Ver second buffer 32). With this configuration, for example, the elongated connected pixel group (LongCC) in the orthogonal direction that intersects with the determined character string, such as being shortened by the length of the intersection, is evaluated according to the area of the intersection with the black rectangle. Is deducted. Thereafter, the character string selection unit 30 re-evaluates the elongated connected pixel group (LongCC) of the next rank in consideration of the deduction due to the black rectangle in the orthogonal buffer (Hor second buffer 31, Ver second buffer 32), Passing point (for example, if the length is above the threshold and the deduction area is 20%
The following is extracted as a character string. This extraction result is sequentially stored in a quadrature buffer for deduction points.
(Hor second buffer 31, Ver second buffer 32). In case of failure, no black rectangle for deduction is written. This makes it possible to unify the original character direction even in a dense area of the character string.

Here, the black rectangle for deduction may be the elongated connected pixel group (LongCC) itself. In addition, it is effective that the order of checking the character strings is not fixed in the initial sort order, and that the remaining elongated connected pixel group (LongCC) is dynamically rearranged each time one character string is determined. . However, in such a case, there is a concern that the processing speed will be slowed down. Therefore, it is preferable to adopt the method in consideration of the processing speed and its effect. FIG. 7 specifically shows a state in which character strings are selected and ranked. Number in figure
(1 to 10) ranks character string candidates from the first to tenth. The alphabetic characters (a to f) indicate the intersection of the vertical candidate and the horizontal candidate in the order of processing. 1st, 3rd, 4th, 9th in portrait, 2nd in landscape, 6
The character string candidates at the 7th and 7th places are selected as character strings without deduction. The horizontal character string at the fourth place corresponding to the postal code 52 in FIG. 3 is deducted only at the intersection of "a", but is selected because its ratio is small. The eighth position of the vertically long character string is selected by deducting only the intersection of "b". The character string “ta” in FIG. 3 which is the tenth place (the elongated connected pixel group (LongCC) 61 in FIG. 5)
“Y”, “Agency”, and “Co.” are not selected because the intersections of “c”, “d”, “e”, and “f” are deducted. as a result,
It is possible to prevent the selection of a pseudo character string that is not an original character string and is arranged in the orthogonal direction.

Depending on the threshold value of the evaluation point of the elongated connected pixel group (LongCC), the intersection of the elongated connected pixel group (LongCC) may belong to both the vertical and horizontal character strings. (Not shown) so that each character size connected pixel group (CharCC) does not belong to a plurality of character strings by estimating to which of the address patterns of the entire mail registered in advance in advance (step 110). Alternatively, the character string selection unit 30 may reconsider the consistency of each character string with respect to vertical and horizontal overlap, and make sure that each character size connected pixel group (CharCC) belongs to only one character string. You can also FIG. 8 shows an extracted character string superimposed on the original image. In the figure, for example, the postal code character string 62 intersects with the address character string 59 and the character string 63, and this image is obtained by combining a postal code frame with a vertically written handwritten address from a character string pattern. It can be estimated that For example, the arrangement of the elongated connected pixel group (LongCC) is such that the fifth-place horizontally long character string as shown in FIG. First and eighth vertical character string groups,
Consider the case where the third, ninth and fourth positions exist. In this case, it is determined that the horizontal character string is a postal code character string, and the vertical character string group is more likely to be an address as it exists on the right side, and is more likely to be an address as it exists on the left side, Perform recognition processing. In addition, when the elongated connected pixel group (LongCC) is arranged in several rows in the horizontal direction, it is determined that there is a high possibility that the postal code appears at the top, and that the address and the address appear in order. , And perform subsequent recognition processing. Thus, in this practical example, the zip code string 62 has priority,
"0""3" from the address character string 59 and the character string 63
The character size connected pixel group (CharCC) such as “5” is excluded. The estimation of the postal code character string from the character string group can be performed using the number of character size concatenated pixel groups (CharCC) in the character string, the position of the character string, knowledge of the address pattern, and the like. For a character string recognized as a postal code character string, recognition processing is performed by matching a handwritten numeral with a dictionary prior to recognition processing of another character string. , A recognition process using a handwritten name dictionary is performed. For the address character string, a recognition process is performed in which a comparison with an address retrieved from a postal code, a recognition process using a handwritten address dictionary, and the like are performed. Here, the character group 60 of “IBM” in FIG. 8 is written horizontally, but since the character string 63 written vertically is selected first, it is deducted and is not selected as a character string.
This is not a single process of extracting a character string, but is to be processed by feedback of recognition in the character string 63 or the like.

As described above, according to the method of extracting a destination character string composed of a handwritten character string in the present embodiment, flexibility is higher than that of another method of logically combining character size connected pixel groups (CharCC). Yes, processing is easy. In addition, a character string is determined from the narrow connected pixel group (LongCC) having a high priority, and the evaluation point of the narrow connected pixel group (LongCC) orthogonal to the character string is reduced. This prevents extraction of the elongated connected pixel group (LongCC) as a character string. As a result, even an address having an irregular arrangement of character strings can sufficiently withstand practical use.

Next, FIG. 1 and FIG.
This will be described with reference to FIGS. FIG. 9 is a flowchart for explaining a general flow of reading the address of a print character. FIGS. 10 to 14 show a second practical example of an original image (FIG. 10) and a character size connected pixel group (CharC
C) extraction state (Fig. 11), mesh image ON
FIG. 12 is a diagram showing extracted address area candidates (FIG. 14) in a state where mesh images are turned on after integration (FIG. 13).

First, an image of a mail is read by the image input device 2 (step 121). The mail 70 as shown in FIG.
2, distribution information 73, display 74 of trademark, etc., contact information 75,
Various types of image information such as background print information 76 are included. The image information is binarized by an appropriate method by the binarization unit 11 (step 122), and the connected pixel group (CC) is detected by the connected pixel group (CC) detection unit 13 and stored in the memory.
(Not shown) (step 123). From the connected pixel group (CC), a character size connected pixel group (CharCC) that is not too large or not too small is a character size connected pixel group (Char
CC) is extracted by the extraction unit 14 (step 124).
Note that, unlike the above-described handwritten character string extraction, the maximum value of the threshold is small (for example, 30 pixels) in reading the address area of the printed character. In FIG. 11, the character size connected pixel group (CharCC) is a black rectangle, and is superimposed on the original image for easy understanding. In this practical example, a lot of character information is extracted, but the characters and patterns of the display 74 such as a trademark and the character 82 of “M” and the character 83 of “2000” in the background print information 76 exceed 30 pixels. Excluded as large letters. Also, the picture 84 is omitted. It should be noted that characters 80 of “river” and characters 81 of “two” in the address information 71 are not extracted.

Next, a corresponding mesh image detecting unit determines which central pixel of the character size connected pixel group (CharCC) belongs to a simplified mesh image which is a rectangular area assumed by the mesh image forming unit 41. The calculation is made at 42, and the mesh is set as a black mesh (step 125). FIG.
Here, the image is divided into meshes 88 in a small area of 24 pixels × 24 pixels. Here, the boundary of the mesh 88 is not displayed. Also,
The “−” character part of the address information 71 described in FIG. 10 is white. In order to interpolate such a case, the following step processing is executed. That is, the mesh image connection detection unit 43 converts a white mesh, which is a black mesh on the top and bottom or left and right, into black in order to integrate even if there is a gap for one line or one character (step 126). The result is shown in FIG. As is apparent from comparison with FIG. 12, it can be understood that even if there is a gap corresponding to one line or one character, it is integrated and converted to black. In this practical example, although it is not necessary to fill the space between lines, this processing is effective in that the space between lines is usually filled by filling the gap, and two separated lines can be integrated. As a result, it is possible to grasp a portion that may be an address area as a block as one connected mesh (CM).

Next, the specific area extracting section 45 regards one mesh as one pixel, and sets a rectangle surrounding eight links as a candidate for a destination area (step 127). The part enclosed by the broken line in FIG. 14 is the part extracted in this step, and the areas 90 to 94 correspond to the candidates for the destination area. Here, in the postal matter 70 shown in FIG. 14, the fee-based payment printing 72, the distribution information 73, the display 74 of the trademark and the like described in FIG.
(CM) is not selected. These are excluded from the candidates for the destination area because, for example, the character string is one line, or a sufficient connected pixel group (CC) does not exist. Next, character extraction is performed in the area,
Each area is evaluated based on various attributes of the attribute information storage unit 44, and is ranked as address candidates (step 1).
28). As described above, these various attributes are contents such as whether the character is not close to the end of the image or whether the character string is appropriate. Thus, in FIG. Candidates, 91 are second, 92 are third, 93 are fourth, and 94 are fifth. Next, the address recognition is performed by the image recognition device 3 in the order of the regions having the highest priority, and if a satisfactory result is obtained, the address recognition is terminated there (step 129). With the flow of these processes, a series of reading of postal addresses in the present embodiment ends.

According to the print character string extraction algorithm according to the present embodiment, a rectangular area including the address area is extracted by a simple algorithm for a mail including the sender area and various patterns in addition to the address area. It becomes possible. That is, since each area candidate has a very small area compared to the whole image and is free from patterns and noises, and since it is possible to predict whether the character string in the area is vertical or horizontal, such area candidates are By performing the address recognition for, the detection accuracy can be improved, the processing can be simplified, and the processing speed can be improved accordingly. For this reason, according to this algorithm, it is possible to extract, with a considerably high probability, a destination area of a print in the area, in particular, a print printed by a printer. Further, even with handwriting, it is possible to accurately extract an address written in a horizontal manner such as a postal mail in Europe and America.

As described above, according to the present embodiment, a case where addresses are collectively written like printed characters, or a case where addresses are written in a horizontal manner like European and American mail even if they are handwritten Is easily processed by the above-described algorithm for extracting printed character strings. For irregular addresses that are mixed with horizontal and vertical writing, such as Japanese handwritten addresses, the aforementioned handwritten character string extraction is performed. A character string can be extracted by applying the algorithm of (1). As a result, it is possible to accurately and quickly extract a character string including an address area without selecting a type of address writing.

[0042]

As described above, according to the present invention,
A complicated integration process in all pixels can be avoided, and a character string such as a handwritten address can be extracted with high accuracy and high speed.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram illustrating an overall configuration of an image processing apparatus according to an embodiment.

FIG. 2 is a flowchart for explaining a general flow of address reading mainly with a handwritten character string in the present embodiment.

FIG. 3 is a diagram showing an original image in a first practical example.

FIG. 4 shows a character size connected pixel group in the first practical example.
It is the figure which showed the extraction state of (CharCC).

FIG. 5 shows a horizontally elongated connected pixel group in the first practical example.
It is the figure which showed the extraction state of (lateral direction LongCC).

FIG. 6 shows a vertically elongated connected pixel group in the first practical example.
FIG. 14 is a diagram illustrating an extraction state of (vertical LongCC).

FIG. 7 is a diagram showing a state in which character strings in the first practical example are selected and ranked.

FIG. 8 is a diagram showing an extracted character string superimposed on an original image in the first practical example.

FIG. 9 is a flowchart for explaining a general flow of reading an address of a print character.

FIG. 10 is a diagram showing an original image in a second practical example.

FIG. 11 is a diagram showing an extraction state of a character size connected pixel group (CharCC) in a second practical example.

FIG. 12 turns on the mesh image in the second practical example.
FIG.

FIG. 13 is a diagram showing a state in which mesh images are turned on after being integrated in the second practical example.

FIG. 14 is a diagram illustrating extracted destination area candidates in a second practical example.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Character string extraction device, 2 ... Image input device, 3 ... Image recognition device, 11 ... Binarization part, 12 ... Reduction part, 13 ... Connection pixel group (CC) detection part, 14 ... Character size connection pixel group ( CharC
C) Extraction unit, 15: threshold storage unit, 16: size determination unit,
21: horizontal extension, 22: vertical extension, 23: Ho
r first buffer, 24... Ver first buffer, 25, 2
6 ... elongated connected pixel group (LongCC) extraction unit, 27 ... horizontal evaluation unit, 28 ... vertical evaluation unit, 29 ... character string threshold storage unit,
30: character string selection unit, 31: Hor second buffer, 32
... Ver second buffer, 40... Print character string extraction unit, 41
... Mesh image forming section, 42... Corresponding mesh image detecting section, 43... Mesh image connection detecting section, 44... Attribute information storing section, 45.

 ──────────────────────────────────────────────────続 き Continued on the front page F term (reference) 3F079 AA02 AA03 CA02 CA03 CA11 CB08 CB29 CB31 CB35 5B029 AA01 CC28 EE14 5B064 AA04 CA08 5L096 AA07 BA08 BA17 EA02 EA03 FA64 FA67 GA15 GA51

Claims (15)

[Claims] 1. A step of extracting a connected pixel group of pixels constituting an input image, and comparing a size of the extracted connected pixel group with a predetermined threshold to extract a connected pixel group within a threshold. And extracting a connected elongated pixel group in the vertical or horizontal direction of the extracted connected pixel group within the threshold, and a character string based on the connected state of the extracted connected elongated pixel group. Extracting a candidate. 2. The step of extracting a connected pixel group within a threshold value includes comparing a vertical direction and / or a horizontal direction of the connected pixel group with a predetermined vertical direction and / or a horizontal direction of an assumed character size. 2. The character string extracting method according to claim 1, wherein the intra-threshold connected pixel group is extracted by: 3. The method according to claim 1, wherein the step of extracting the elongated connected pixel group comprises extracting the elongated connected pixel group by changing a size of the extended connected pixel group based on the size of the extracted intra-threshold connected pixel group. 2. The character string extraction method according to claim 1, wherein: 4. The method according to claim 1, wherein the step of extracting the group of elongated connected pixels includes the step of contracting the horizontal direction when the group of connected pixels within the threshold value is elongated in the vertical direction, and shrinking the vertical direction when the group of elongated connected pixels is elongated in the horizontal direction. 2. A connected pixel group is extracted.
The described character string extraction method. 5. The step of extracting the elongated connected pixel group is characterized in that the elongated connected pixel group is extracted by changing an amount of contraction based on the size of the extracted connected pixel group within the threshold. The character string extraction method according to claim 4. 6. Assuming a mesh image obtained by dividing the image into a predetermined size, turning on the mesh image corresponding to the extracted connected pixel group within the threshold, and concatenating the mesh images turned on. 6. The method according to claim 5, further comprising: extracting a specific area from the state. 7. The mesh image is turned on when the image is a matched character string such as a print character as an address character string.
7. The method according to claim 6, further comprising the step of: extracting the elongated connected pixel group when the image is an inconsistent character string such as Japanese handwritten characters as an address character string. The described character string extraction method. 8. A pixel group having a predetermined size is extracted from an image including an input handwritten character, and a string of a character string in the expanded direction is obtained by extending one direction of the pixel group and contracting the other direction. A handwritten character string extraction method, comprising: detecting a string state of a plurality of assumed character strings; and estimating an address pattern from the detected string state of the character string. 9. The detected state of the string of the character string includes the length of the string and the unevenness state of the string, and the order in which the character recognition processing is performed is the one in which the length of the string is long or 9. The method for extracting a handwritten character string according to claim 8, wherein the order is raised for / and the one with the small unevenness state. 10. The string of the assumed character string is extended by adding a predetermined ratio to the size of the pixel group with respect to the vertical or horizontal one of the pixel groups, 9. The string of the character string is assumed based on a connection state of a plurality of pixel groups that are reduced, reduced, expanded, and reduced by reducing a predetermined ratio to the size of the pixel group with respect to the vertical. Extraction method of handwritten character string of description. 11. A connected pixel group detecting section for detecting a connected pixel group formed from black connected pixels in a binarized image, and among the connected pixel groups detected by the connected pixel group detecting section. A character size connected pixel group extraction unit that extracts a character size connected pixel group that is not too large and not too small, and the character size connected pixel group extracted by the character size connected pixel group extraction unit, in the assumed character string direction. Is expanded and contracted in the direction orthogonal to the assumed character string direction to expand and contract the character size connected pixel group, and a plurality of the connected pixel groups expanded and contracted by the connected pixel group expandable unit. An elongated connected pixel group extraction unit for extracting an elongated connected pixel group by combining in the assumed character string direction; and performing image recognition based on the elongated connected pixel group extracted by the elongated connected pixel group extraction unit. Character string extraction apparatus characterized by comprising a string selection unit for selecting a can string. 12. The character size connected pixel group extracting unit,
12. The character string extraction device according to claim 11, wherein the character size connection pixel group is extracted based on a threshold value in consideration of a size of a handwritten character written on a postal matter. 13. An image input unit for inputting an image, a pixel group extraction unit for extracting a pixel group having a predetermined size from the image input by the image input unit, and an image extracted by the pixel group extraction unit. A vertically-extended pixel group generating means for defining a vertical direction and a horizontal direction with respect to the pixel group, and generating a vertically-extended pixel group in which the pixel group is expanded and contracted in the horizontal direction in the vertical direction; A vertical direction and a horizontal direction are defined with respect to the pixel group extracted by the pixel group extracting unit, and a horizontal expanded pixel group is generated in which the pixel group is expanded in the horizontal direction and contracted in the vertical direction. A horizontal extension pixel group generation unit; a plurality of the vertical extension pixel groups generated by the vertical extension pixel group generation unit; and a plurality of the horizontal extension pixels generated by the horizontal extension pixel group generation unit. The,
Elongate connected pixel group extraction means for extracting an elongate connected pixel group connected in the extended direction, and a character string for extracting a character string candidate based on the elongate connected pixel group extracted by the elongate connected pixel group extraction means An image processing apparatus comprising: a candidate extracting unit; and an image recognizing unit that performs image recognition on a character string candidate extracted by the character string candidate extracting unit. 14. The character string candidate extraction unit outputs a plurality of the elongated connected pixel groups extracted by the elongated connected pixel group extraction unit with a higher priority to those having a longer length. The image processing apparatus according to claim 13, wherein: 15. The character string candidate extraction means determines a priority by lowering an evaluation point for a pixel group at which the vertical extension pixel group and the horizontal extension pixel group intersect. The image processing device according to claim 14.